Claude Sonnet 5.5: Cache Reads Halved, Set between_tools First
Claude Sonnet 5.5 was released on September 28, 2026. Input and output pricing is the same as Sonnet 5, still $2 / $10 per million tokens. The change is in cache reads, which drop from $0.20 to $0.10. If you have been running Sonnet with thinking turned off, switch to the new between_tools setting before you migrate. Everything below is current as of October 2026 and comes from Anthropic's official pages; defer to them for details.
What changed from Sonnet 5 to Sonnet 5.5?
According to Anthropic's release page, the model ID is claude-sonnet-5-5. The main differences:
| Item | Sonnet 5 | Sonnet 5.5 |
|---|---|---|
| Input / output price (per million tokens) | $2 / $10 | $2 / $10 |
| Cache read (per million tokens) | $0.20 | $0.10 |
| Output speed | Baseline | 30%+ faster |
| Terminal-Bench 4.0 | 10.3% | 70.6% |
Anthropic says output is more than 30% faster than Sonnet 5, and that in its own testing the cost per task is up to 30% lower than Sonnet 5. Those are results on Anthropic's task set; whether your workload saves that much is something you have to measure. The jump on Terminal-Bench from 10.3% to 70.6% is large, so re-test on your own codebase rather than taking it at face value.
The price is unchanged, so why might the bill go down?
Two reasons: the cache read price is halved, and the model uses fewer tokens to finish the same task. The changelog for Claude Code 2.1.296 (October 9) also notes that cost estimates in /cost, the status line, --max-budget-usd, and the SDK now use the Sonnet 5.5 cache read price of $0.10 per million tokens (previously $0.20).
Workloads with a high share of cache reads benefit most, such as long system prompts or agents that repeatedly read the same set of files. For how to mark cache breakpoints and verify hits, see How to use Claude prompt caching to save money.
How should users who run with thinking off migrate?
The official release page says that anyone running Sonnet with thinking turned off needs to switch to the new between_tools setting first, which keeps the "no thinking at the start" behavior, and only then move to Sonnet 5.5. For the exact parameter syntax and values, see the migration guide linked from the release page. This article does not guess on Anthropic's behalf.
Pre-migration checklist:
- Search your code and config for every place thinking is currently disabled (SDK calls, gateway config, environment variables).
- Following the official migration guide, change those places to
between_tools, and run it once on the old model first. - Once behavior on the old model looks right, change the model ID to
claude-sonnet-5-5. - Use a fixed set of real requests to compare latency, token usage, and output quality before and after.
- Check your bill separately for cache read charges to confirm they are billed at the new rate.
Order matters. Change the setting first, then swap the model, so that if something breaks you can tell which step caused it.
What is the default effort level?
The official defaults: Medium in Claude Code and the Claude app, High on the Claude Platform. The same task can get different amounts of thinking in Claude Code versus a direct API call, so align effort when you compare latency and cost.
What happens with high-risk cybersecurity requests?
Anthropic says higher-risk cybersecurity requests fall back to Sonnet 5, and this is visible to the user. Everyday vulnerability finding and fixing is not affected. Teams doing security research should keep this in mind and log which model actually responded in their request logs.
Sonnet 5.5 or Opus 5.5?
For benchmark comparisons between the two, see the official release page. Those are evaluations Anthropic chose, and they do not represent your tasks. For Opus 5.5 pricing and configuration changes, see Claude Opus 5.5 release. For current Opus and Haiku prices, check Anthropic's official pricing page; this article does not repeat them.
A practical approach:
- For everyday coding, review, and batch jobs, default to Sonnet 5.5 and evaluate on your own samples.
- For complex tasks that Sonnet 5.5 keeps failing, try Opus 5.5 and record whether it actually helps.
- For simple tasks that are sensitive to latency and cost, consider a lighter tier, using the official pricing page as the source of truth.
How do I keep watching after launch?
In the first weeks after a new model ships, availability, latency, and pricing can all shift. Ways to track them:
- Use model monitoring at /models to see listing, delisting, and price-change records.
- Use site changes at /changes to follow additions and removals on Anthropic's pages. For the method, see How to find out right away when an LLM API price changes.
- When you hit errors or slowness, check Anthropic API status at /status/anthropic first, then read Three new Anthropic models in September and October: how to read availability and latency after launch.
If you use Claude Code through a custom endpoint, see Configuring custom API endpoints for Claude Code, Codex, and OpenCode, and confirm that your gateway already supports the new model ID.
